Claude Code source code LEAKED
TL;DR
Anthropic accidentally exposed Claude Code’s client-side guts at massive scale — Wes says a leaked 60MB source map was quickly reconstructed into roughly 200,000 TypeScript files and 600,000 lines of code, with one GitHub mirror hitting about 32,000 stars and 44,000 forks before takedowns landed.
The leak revealed Anthropic’s near-term roadmap, not model weights — hidden feature flags point to real but unshipped capabilities like background agents, persistent memory, browser control, voice mode, and multi-agent coordination, while Wes stresses that no training secrets, customer data, or API keys were exposed.
Claude appears to be moving toward autonomous, always-on coding workflows — standout features include “Chyros,” a background GitHub-monitoring agent, “AutoDream,” which compresses and consolidates context like REM sleep, and “Ultra Plan,” a dedicated 30-minute remote planning session run by a more expensive deep-planning model.
The biggest subtext is legal: AI may make ‘clean room’ cloning cheap and fast — Wes highlights a fork that allegedly translated Anthropic’s TypeScript implementation into Python using OpenAI’s Codex, arguing this could blur the line between stealing code and merely reproducing functionality.
A lot of the roadmap looks like a response to OpenClaw’s momentum — Wes repeatedly connects hidden features such as persistent memory, orchestration, scheduling, and second-opinion commands to the popularity of OpenClaw and says the leak shows how seriously labs are reacting.
The weirdest surprise is a fully built Tamagotchi-style pet system inside Claude Code — Reddit reportedly found a
/buddycommand, 18 pet species including duck, capibara, dragon, ghost, and axolotl, plus rarity tiers, hats, and stats like debugging, chaos, and snark.
The Breakdown
The leak wasn’t a rumor — it was Claude Code’s source map blowing open
Wes opens in full alarm-bell mode: this wasn’t the earlier Mythos leak, but a fresh Anthropic mistake that exposed Claude Code’s internals through an accidentally included 60MB source map. The internet moved instantly — archiving, cloning, and forking it “within hours,” with Wes citing tens of thousands of copies and a mirror that ballooned to 32,000 stars and 44,000 forks.
Why this matters beyond Anthropic: AI may be flattening copyright boundaries
He pivots from the leak itself to the legal bomb underneath it: copying source code is one thing, copying functionality is usually another. The twist here is that someone reportedly converted the leaked TypeScript codebase into Python using OpenAI’s Codex in a matter of hours, which Wes frames as AI-powered “clean room engineering” becoming cheap, fast, and legally murky.
The roadmap hiding behind feature flags
The juicy part of the leak, to Wes, is all the stuff Anthropic already built but hasn’t shipped. He says the code references the leaked Mythos model codenamed Capibara and shows a pile of disabled features that effectively expose where Claude Code is headed next.
Chyros, AutoDream, and the push toward agents that work while you sleep
One standout is “Chyros,” a background coding agent that monitors GitHub repos, keeps working without user input, and can be pinged from anywhere. Another is “AutoDream,” which Wes compares to REM sleep: when nothing else is happening, it reviews past context and consolidates memory instead of just letting everything pile up forever.
Ultra Plan and the rise of multi-agent orchestration
Wes also points to “Ultra Plan,” a separate 30-minute remote planning session that uses a more expensive deep-planning model before execution begins. Alongside it is “coordinator mode,” Anthropic’s more corporate-sounding version of the old “agent swarm” idea — one Claude orchestrating multiple worker agents with their own scratchpads and tools.
Browser control, persistent memory, and OpenClaw’s fingerprints everywhere
The leak suggests real browser control, cron-style scheduling, and persistent memory across sessions are already in the product, just not exposed publicly. Wes’s take is that a lot of this smells like a direct response to OpenClaw’s success — especially features like heartbeat-style memory, orchestration, and commands such as /advisor, /goodclaude, /bughunter, and /teleport.
The Tamagotchi detour is absurdly detailed and very real
Then the video takes a hard left into what sounds like an April Fool’s joke but apparently lives in the code: a virtual pet system. Wes relays Reddit’s findings of a /buddy command, 18 species, rarity tiers up to a 1% legendary drop, hats, shiny variants, and pet stats named debugging, chaos, and snark — all delivered with his disbelief that someone may have rushed a build for a Tamagotchi Easter egg.
Claude may be reading your mood — and there’s still more to uncover
One detail that caught Wes off guard is code suggesting Claude monitors user language to detect frustration, basically asking: is this person about to lose patience? He also flags an unexplained “undercover mode,” 187 different spinner verbs, and references to X402-style agentic crypto payments, ending on the bigger point: Anthropic’s roadmap is now public, competitors are definitely reading it, and this is only the start of what people will dig out.